zhtw-en

English

This model translates Traditional Chinese sentences into English, with a focus on understanding Taiwanese-style Traditional Chinese and producing more accurate English translations.

This model is a fine-tuned version of Helsinki-NLP/opus-mt-zh-en on the zetavg/coct-en-zh-tw-translations-twp-300k dataset.

It achieves the following results on the evaluation set:

• Loss: 2.4350
• Num Input Tokens Seen: 55653732

Intended Uses & Limitations

Intended Use Cases

• Translating single sentences from Chinese to English.
• Applications requiring understanding of the Chinese language as spoken in Taiwan.

Limitations

• Designed for single-sentence translation so will not perform well on longer texts without pre-processing
• Sometimes hallucinates or omits information, especially with short or long inputs
• Further fine-tuning will address this

Training and Evaluation Data

This model was trained and evaluated on the Corpus of Contemporary Taiwanese Mandarin (COCT) translations dataset.

• Training Data: 80% of the COCT dataset
• Validation Data: 20% of the COCT dataset

Chinese

該模型旨在將繁體中文翻譯成英文，重點是理解台灣風格的繁體中文並產生更準確的英文翻譯。

模型基於 Helsinki-NLP/opus-mt-zh-en 並在 zetavg/coct-en-zh-tw-translations-twp-300k 資料集上進行微調。

在評估集上，模型取得了以下結果：

• 損失：2.4350
• 處理的輸入標記數量：55,653,732

預期用途與限制

預期用途

• 將單一中文句子翻譯為英文。
• 適用於需要理解台灣中文的應用程式。

限制

• 本模型專為單句翻譯設計，因此在處理較長文本時可能表現不佳，若未經預處理。
• 在某些情況下，模型可能會產生幻覺或遺漏信息，特別是在輸入過短或過長的情況下。
• 進一步的微調將有助於改善這些問題。

訓練與評估數據

該模型使用 當代台灣普通話語料庫 (COCT) 資料集進行訓練和評估。

• 訓練資料：COCT 資料集的 80%
• 驗證資料：COCT 資料集的 20%

Example

from transformers import pipeline

model_checkpoint = "agentlans/zhtw-en"
translator = pipeline("translation", model=model_checkpoint)

# From Chinese Wikipedia's article of the day (摘自中文維基百科的今日文章)
translator("《阿奇大戰鐵血戰士》是2015年4至7月黑馬漫畫和阿奇漫畫在美國發行的四期限量連環漫畫圖書，由亞歷克斯·德坎皮創作，費爾南多·魯伊斯繪圖，屬跨公司跨界作品。")[0]['translation_text']

# Output (輸出)
# Acer's Iron Blood Fighter is a four-year series of comic books published in the United States by Black Horse and Ah Chi comics from April to July of that year. The book was created by Alexander d'Campie and painted by Philnanto Ruiz. It is a cross-firm work.

# Compare with my own gold standard translation: (與我自己的黃金標準翻譯比較:)
# "Archie vs. Predator" is a limited four-issue comic book series published by Black Horse and Archie Comics from April to July 2015. It was created by Alex de Campi and drawn by Fernando Ruiz. It's a crossover work.

Training Procedure

Training Hyperparameters

The following hyperparameters were used during training:

• Learning Rate: 5e-05
• Train Batch Size: 8
• Eval Batch Size: 8
• Seed: 42
• Optimizer: adamw_torch with betas=(0.9,0.999) and epsilon=1e-08
• LR Scheduler Type: linear
• Number of Epochs: 3.0

Training Results

Click here to see the training and validation losses

Training Loss	Epoch	Step	Validation Loss	Input Tokens Seen
3.2254	0.0804	2500	2.9105	1493088
3.0946	0.1608	5000	2.8305	2990968
3.0473	0.2412	7500	2.7737	4477792
2.9633	0.3216	10000	2.7307	5967560
2.9355	0.4020	12500	2.6843	7463192
2.9076	0.4824	15000	2.6587	8950264
2.8714	0.5628	17500	2.6304	10443344
2.8716	0.6433	20000	2.6025	11951096
2.7989	0.7237	22500	2.5822	13432464
2.7941	0.8041	25000	2.5630	14919424
2.7692	0.8845	27500	2.5497	16415080
2.757	0.9649	30000	2.5388	17897832
2.7024	1.0453	32500	2.6006	19384812
2.7248	1.1257	35000	2.6042	20876844
2.6764	1.2061	37500	2.5923	22372340
2.6854	1.2865	40000	2.5793	23866100
2.683	1.3669	42500	2.5722	25348084
2.6871	1.4473	45000	2.5538	26854100
2.6551	1.5277	47500	2.5443	28332612
2.661	1.6081	50000	2.5278	29822156
2.6497	1.6885	52500	2.5266	31319476
2.6281	1.7689	55000	2.5116	32813220
2.6067	1.8494	57500	2.5047	34298052
2.6112	1.9298	60000	2.4935	35783604
2.5207	2.0102	62500	2.4946	37281092
2.4799	2.0906	65000	2.4916	38768588
2.4727	2.1710	67500	2.4866	40252972
2.4719	2.2514	70000	2.4760	41746300
2.4738	2.3318	72500	2.4713	43241188
2.4629	2.4122	75000	2.4630	44730244
2.4524	2.4926	77500	2.4575	46231060
2.435	2.5730	80000	2.4553	47718964
2.4621	2.6534	82500	2.4475	49209724
2.4492	2.7338	85000	2.4440	50712980
2.4536	2.8142	87500	2.4394	52204380
2.4148	2.8946	90000	2.4360	53695620
2.4243	2.9750	92500	2.4350	55190020

Framework Versions

• Transformers 4.48.1
• Pytorch 2.3.0+cu121
• Datasets 3.2.0
• Tokenizers 0.21.0